Nirajan Paudel

I work on natural language processing for languages that the field has mostly skipped. Nearly every technique that makes large language models fast, cheap, or steerable was developed and measured in English — and a surprising number of them quietly stop working somewhere else.

Our recent work shows this concretely for speculative decoding. The speed-ups everyone quotes hold up in English and then thin out, language by language, as training resources disappear. The interesting part isn't that it degrades; it's why the obvious fixes don't repair it.

I grew up in Nepal. Nepali has around 32 million speakers and almost no NLP infrastructure, which makes it a recurring test case in my work rather than a footnote. I also build things — production ML systems, developer tools, infrastructure — because the engineering keeps the research honest.